Man Zhou is currently working toward the PhD degree in University of Science and Technology of China, Hefei, China. His research interests include image/video processing, computer vision
Xueyang Fu received the PhD degree in signal and information processing from Xiamen University, Xiamen, China, in 2018. He was a Visiting Scholar with Columbia University, New York, USA, sponsored by the China Scholarship Council, from 2016 to 2017. He is currently an Associate Researcher with the Department of Automation, University of Science and Technology of China, Hefei, China. His research interests include machine learning and image processing
Visual features with high potential for generalization are critical for computer vision applications. In addition to the computational overhead associated with layer-by-layer feature stacking to produce multi-scale feature maps, existing approaches also incur high computational costs. To address this issue, we present a compact and efficient scale-in-scale convolution operator called SIS by incorporating an efficient progressive multi-scale architecture into a standard convolution operator. More precisely, the suggested operator uses the channel transform-divide-and-conquer technique to optimize conventional channel-wise computing, thereby lowering the computational cost while simultaneously expanding the receptive fields within a single convolution layer. Moreover, the proposed SIS operator incorporates weight-sharing with split-and-interact and recur-and-fuse mechanisms for enhanced variant design. The suggested SIS series is easily pluggable into any promising convolutional backbone, such as the well-known ResNet and Res2Net. Furthermore, we incorporated the proposed SIS operator series into 29-layer, 50-layer, and 101-layer ResNet as well as Res2Net variants and evaluated these modified models on the widely used CIFAR, PASCAL VOC, and COCO2017 benchmark datasets, where they consistently outperformed state-of-the-art models on a variety of major vision tasks, including image classification, key point estimation, semantic segmentation, and object detection.
Graphical Abstract
We design a series of transformation mechanism to expand the module (a) and Res2Net module (b) into multi-scale module operators (c, d, e, f, g).
Abstract
Visual features with high potential for generalization are critical for computer vision applications. In addition to the computational overhead associated with layer-by-layer feature stacking to produce multi-scale feature maps, existing approaches also incur high computational costs. To address this issue, we present a compact and efficient scale-in-scale convolution operator called SIS by incorporating an efficient progressive multi-scale architecture into a standard convolution operator. More precisely, the suggested operator uses the channel transform-divide-and-conquer technique to optimize conventional channel-wise computing, thereby lowering the computational cost while simultaneously expanding the receptive fields within a single convolution layer. Moreover, the proposed SIS operator incorporates weight-sharing with split-and-interact and recur-and-fuse mechanisms for enhanced variant design. The suggested SIS series is easily pluggable into any promising convolutional backbone, such as the well-known ResNet and Res2Net. Furthermore, we incorporated the proposed SIS operator series into 29-layer, 50-layer, and 101-layer ResNet as well as Res2Net variants and evaluated these modified models on the widely used CIFAR, PASCAL VOC, and COCO2017 benchmark datasets, where they consistently outperformed state-of-the-art models on a variety of major vision tasks, including image classification, key point estimation, semantic segmentation, and object detection.
Public Summary
A more lightweight and representative scale-in-scale operator, namely SIS is proposed. The operator can be plugged into any promising backbone to replace regular convolution operator.
To be more efficient, an improved SIS series is proposed. The series maintains promising results within nearly half parameters.
Extensive experiments demonstrate the superior performance of our SIS operator series compared with state-of-the-art methods on image classification, object detection, key points estimation and semantic segmentation.
Advance in decision-making research indicates that behavior is governed by at least two separable control systems[1,2]: System 1 is automatic and error prone, while System 2 is cognitively demanding and reliable[3]. Due to the complementary nature of these two systems in terms of cost and benefit (resulting from behavior), they apply to different task scenarios[4]. As a control mechanism for monitoring and adjusting cognitive control, metacontrol can dynamically regulate these two control systems based on a cost‒benefit analysis of the current context[5]. Recent decision-making research has discovered a noticeable weakening of metacontrol in aging (i.e., changes in cost and benefit do not impact system-level control)[6]. However, the underlying mechanism behind metacognitive inefficiency remains elusive.
The hybrid reinforcement learning model provides a quantitative tool for evaluating metacontrol[4]. The computational model associates System 1 and System 2 with model-free and model-based reinforcement learning methods, in which the model-free method directly establishes the direct association between actions and rewards, while the model-based method plans based on learned structural knowledge (the mapping between actions and states)[7]. Assuming that the computational results of the two systems are mixed by linear weighting, the model-based weights in the hybrid reinforcement learning model represent the relative contribution of the model-based system to the decision. Metacontrol studies usually set independent model-based weights in different task contexts and analyze the mechanism of metacontrol by observing the influence of experimental conditions on the weights[6].
Theoretical work describes the metacontrol process as follows: participants estimate the expected rewards of each system in the current situation, then weigh it against the cost of model-based control to find a balance point for control, and finally regulate the intensity of model-based control[2,8]. This description has received considerable empirical support. For example, when the stake (reward magnitude) is high, people tend to increase their cognitive investment in model-based control[4,9]. However, when planning complexity increases or the structure turns variable, people tend to decrease their model-based control[2].
Older adults exhibit weakened metacontrol compared to younger adults[6]. The inefficiency of metacontrol may be due to age-related differences in learning and maintaining task structural representation[10–12]. Obviously, metacontrol cannot work properly even if the structural knowledge is poorly represented. Alternatively, metacontrol inefficiency might stem from reward insensitivity or reduced willingness or motivation to perform cost‒benefit analysis[6]. This view is consistent with recent research that suggests the existence of age-related deterioration of stimulus-reward processing[13]. However, there is also accumulating evidence that older people are just as sensitive to rewards as younger people. For instance, functional magnetic resonance imaging studies using incentive delay tasks suggest fewer age-related changes in reward network responses during anticipation and outcome processing[14]. In addition, it has been observed that participants exhibit more vigorous movement responses to obtain rewards when the average reward rate is higher, and this response is similar between older and younger groups[15].
The present study investigates the cognitive and motivational factors underlying metacontrol inefficiency by combining pupillometry and behavior experiments. Similar to Bolenz et al.[6], we used a sequential decision-making task that manipulated reward stake (low-stake vs. high-stake) and task structure variability (stable transition vs. variable transition). In addition, there were two methodological improvements. First, pupillometry recording was used to measure motivation to allocate cognitive effort[16,17]. Second, we used a structured questionnaire to ask participants’ knowledge about the task background and task structure. The novel design of the experiment enabled us to systematically examine the motivational and cognitive differences between older and younger adults. Consistent with previous research[6], we predicted that older adults should exhibit weaker metacontrol than younger adults, manifested as the insignificant influence of stake and transition conditions on model-based weights in behavioral analysis. However, we predicted that older adults should be sensitive to stake levels and feedback but insensitive to structural needs in pupillometry analysis. In addition, we expected that the intactness of structure knowledge reflected in the subjective oral reports determined whether older adults could show comparable metacontrol efficiency as younger adults. Last, we used computational simulation to demonstrate that poor structure knowledge representation impaired metacontrol.
2.
Materials and methods
2.1
Participants
We recruited 32 healthy younger and 55 healthy older adults. Before the appointment, participants were telephone screened for potential exclusion due to psychiatric or neurological illness. All subjects had normal or corrected-to-normal vision and abstained from caffeine in the hour immediately preceding their participation. Seven participants were excluded for failing to complete the task, leaving a final sample of 29 younger adults (age range 20–44 years, mean age 30.14 years, 14 female) and 51 older adults (age range 60–83 years, mean age 68.23 years, 26 female). The sample size was comparable to that in a previous study[18].
Pupil sizes were measured using a desktop eye-tracking platform (PupilEXT) with a sampling rate of 100 Hz[19]. Pupil diameter was determined by employing the Pure algorithm[20]. The stimulus materials utilized in the previous study[6] were employed. The Ethics Committee of Hefei Institutes of Physical Science, Chinese Academy of Sciences, approved the research protocol (No. YXLL-2023-32). All participants provided written informed consent and received monetary incentives after completion of the study.
2.2
Experimental procedure
The experimental procedure included a practice session, quiz A, a formal session and quiz B. In each trial, the participants viewed a cue followed by two-step choices and then feedback indicating how much reward they received. Three hundred twenty trials were equally divided by the combination of low- or high-stake and stable- or variable-transition.
Each trial began with one of two first-stage states, each containing two alternatives (Fig. 1a). All four first-stage choices deterministically led to one of the two second-stage states. In the second stage, participants had to press the space bar to earn scalar points generated by the random walk (ranging from 0 to 9).
Figure
1.
Behavioral task. (a) State transition structure of the task. Each trial starts with a random first-stage state. Given the transition structure, each first-stage choice leads deterministically to the second-stage state. Each second-stage choice is associated with a drifting scalar reward. (b) The stake manipulation (top). One of the high-stake or low-stake tips is randomly presented at the beginning of the trial, which means that the actual benefit of the trial is several times the score shown in the feedback. Transition manipulation (bottom). The task transition did not change in stable-transition blocks, and the task transition structure changed irregularly in variable-transition blocks.
The experiment comprised 320 trials organized into four blocks of 80 trials each: two stable transition blocks and two variable transition blocks. These blocks alternated in sequence. In half of the trials, a high-stake cue was randomly assigned, while the remaining half received a low-stake cue (Fig. 1b). At the onset of each trial, the stake cue provided information to the participant regarding the trial’s stakes condition. In other words, these cues determined how rewards were converted into points. In low-stake trials, the obtained reward was multiplied by a factor of 1, whereas in high-stake trials, the reward was multiplied by a factor of 5. The transition structure remained the same throughout the block in stable transition blocks. In contrast, during variable transition blocks, changes in the trial structure would occur every 6 to 14 trials.
The whole process, including receiving instructions, performing practice, taking two quizzes and a test session, and the final payment, took between 80 and 110 min for each subject. The experimental stimulus material was partly drawn from a shared repository from previous studies[21], and the paradigm is similar to Bolenz et al[6]. The experiment was conducted using jsPsych[22] and executed on a standard PC. All participants controlled the experiment using a computer keyboard. The left and right spaceships were chosen using the F and J keys, respectively, while the space bar was used for the second-stage response. The response time limit for both the first-stage and second-stage phases was set at 2 s.
In the practice session, participants were repeatedly instructed to choose one spaceship in the first stage, transition to one planet in the second stage, and eventually receive the random-walked reward. Participants were instructed extensively about the reward distribution, the transition structure, and the stakes manipulation. Participants must choose the correct path to a specified planet ten times in a row and answer the correct score they received under a specified stake condition ten times in a row. Participants also had to complete 40 practice trials that were divided into two blocks within stable and variable transitions.
Before the formal testing phase, we used quiz A to test participants’ understanding of the task rules. Quiz A includes:
(i) “How many spaceships are in the task?” The options include “2” and “4”. The correct answer is “4”.
(ii) “How many planets are in the task?” The options include “2” and “4”. The correct answer is “2”.
(iii) “If you got 4 treasures on the purple planet in the last trial, what is the range of the number of treasures you are most likely to get on the purple planet in this trial?” The options include “0–2”, “3–5” and “7–9”. The correct answer is “3–5”.
(iv) “The treasure yield on the red planet is increasing. What about the treasure yield on the purple planet?” The options include “increasing”, “decreasing”, “remain the same”, and “All of the above could happen.” The correct answer is “All of the above could happen.”
(v) “If we saw spaceship 1 in the first stage, what would the other ship be?” The options include “spaceship 2”, “spaceship 3”, “spaceship 4”, and “All of the above could happen.” The correct answer is “spaceship 2”.
(vi) “If spaceship 3 goes to the purple planet, what planet does spaceship 4 go to?” The options include “red planet”, “purple planet”, and “All of the above could happen.” The correct answer is “red planet”.
(vii) “If spaceship 1 always goes to the red planet, this turn suddenly brings you to the purple planet, what’s the reason?” The options include “I pressed the wrong key.” and “The transition structure has just changed.” The correct answer is “The transition structure has just changed.”
Participants who scored less than 5 points on quiz A dropped out of the experiment, and for the rest of the participants, the experimenter gave them an oral explanation of the incorrectly answered questions from a prewritten manuscript.
The left and right spaceships were chosen using the F and J keys, respectively, and the second-stage response was executed by pressing the space bar. If participants failed to respond within the 2-second time limit for both the first-stage and second-stage states, the experiment advanced to the next trial. If responses were made within the allotted time, the selected option remained highlighted for the remainder of the response window.
After the test session, the participants were asked to complete quiz B. The quiz contained the entire content of the previous quiz A and the following questions:
(viii) “What is the main basis for choosing the spaceship?” The options include “intuition”, “ reasoning”, and “both”.
(ix) “Will you be more serious when the price sign is five times higher?” The options include “yes” and “no”.
(x) “Do you always remember that more gems can be exchanged for more cash?” The options include “yes” and “no”.
(xi) “How long do you need to redefine the route when the route schedule changes?” The options include “Easy: 1–3 trials”, “Medium: more than 4 trials”, and “Hard: It is hard to figure it out.”
2.3
Preprocessing of pupillometric data
Preprocessing of the pupil data was conducted based on previous studies[16] and involved several steps. First, eye blink artifacts were removed by excluding data from 100 ms before the blink to 150 ms after. Pupil sizes smaller than 1 mm were also excluded, and data points that exceeded the inner fences were removed. In cases of missing data, linear interpolation was applied up to 500 ms. To eliminate slow drift below 0.012 Hz, a high-pass Butterworth filter was utilized on the pupil diameter measurements.
Three distinct processing modes are subsequently distinguished. When analyzing the pupil response induced by the stake, the pupil diameter values were standardized (Z scored) within blocks to ensure comparability across different blocks. Following this, baseline correction was performed by subtracting the mean diameter of a 200 ms baseline period preceding the trial onset. Alternatively, when analyzing the pupil response induced by feedback, baseline correction was conducted by subtracting the mean diameter of a 200 ms baseline period before the feedback phase onset. When comparing the pupil baseline of different blocks, Z-scoring within the blocks was not applied. Due to the fixed distance between the camera and the headrest, the pupil diameters recorded in pixels by PupilEXT were converted into actual diameters using a fixed coefficient.
2.4
Computational model
We implemented an established hybrid reinforcement learning algorithm in our study[6,21,23]. This algorithm combines elements from both model-based learning methods and the model-free SARSA(λ) algorithm[24] to learn action values.
For each trial t, δ1 and δ2, the reward prediction errors of the first stage and second stage are calculated as:
δ1=QMF(s2,t,a2,t)−QMF(s1,t,a1,t)，
(1)
δ2=rt−QMF(s2,t,a2,t),
(2)
where s1,t and s2,t are the first stage and second stage at trial t, respectively. a1,t and a2,t are the actions in the first stage and second stage at trial t, respectively. (s1,t, a1,t) and (s2,t, a2,t) is the state-action pair in the first stage and second stage at trial t, respectively. rt is the received immediate reward at trial t.
The update of the model-free state-action value QMF at the first stages and second stages is performed as:
QMF(s1,t,a1,t)=QMF(s1,t,a1,t)+αδ1+αλδ2,
(3)
QMF(s2,t,a2,t)=QMF(s2,t,a2,t)+αδ2,
(4)
where α is the learning rate (0 ⩽α⩽ 1), and λ is the eligibility trace decay (0⩽λ⩽1) that modulates the extent of the reward prediction error in the second stage influencing the first-stage state-action value.
The model-based system learns a transition probability matrix T(s2∣s1,t,a1,t) representing the probability of transitioning to the second-stage state s2 after selecting an action a1 in the first-stage state s1. The transition probability matrix can be set to be an invariant matrix or updated by state prediction errors. Please refer to Daw et al.[23] and Bolenz et al.[6] for detailed model information.
The model-based state-action value at the first stage is given by:
QMB(s1,t,a1,t)=∑s2T(s2∣s1,t,a1,t)QMF(s2,t,a2,t).
(5)
Outside of the model-based and model-free systems, the model-free and model-based state-action values are combined according to the model-based weight ω (0 ⩽ω⩽ 1):
Qnet(s1,a1)=ωQMB(s1,a1)+(1−ω)QMF(s1,a1).
(6)
The choice probabilities were calculated as follows:
where β is the inverse SoftMax temperature. The choice stickiness parameter π captures the stimulus preference, and the response stickiness ρ captures the key preference. rep(a1) is equal to 1 when choosing the same stimulus as the last trial and 0 otherwise. resp(a1) is equal to 1 when choosing the same key as the last trial and 0 otherwise.
2.5
Statistical analyses
Paired-sample Student’s t tests (two-tailed) were used for comparisons. Hierarchical Bayesian regression analyses were performed in R Studio using the brms package[25]. Effect coding was employed for categorical variables (younger adults = −1, older adults = 1, low-stake = −1, high-stake = −1, stable-transitions = −1, variable-transitions = −1), allowing for the interpretation of regression coefficients as main effects. Each reported regression coefficient is characterized by the mean of its marginal posterior distribution and the 95% credible interval. The interval is computed as the [0.025,0.975] percentile range, and it can be understood as containing the parameter of interest with 95% probability.
3.
Results
We conducted a metacontrol task (Fig. 1) in younger and older adults. The task involved presenting stake cues randomly at the onset of a single sequential decision-making trial, with the cues indicating either a high or low reward magnitude. Additionally, the task manipulated the level of structural variability at the block level by either keeping it stable or making it variable. For each participant, we estimated model-based weights in four different scenario demands. We compared the estimated model-based weights parameter with the hierarchical regression analysis in all participants and discovered that aging had reduced model-based weights (βage group = −0.06, 95% CI = [−0.08, −0.04]).
3.1
Reduced metacontrol in older adults
We separately compared the model-based weights with an effect-coded hierarchical regression analysis in younger and older adults (Fig. 2). For younger adults, the results indicated a main effect of stakes condition (βstake = 0.04, 95% CI = [0.01, 0.07]), indicating an elevation in model-based control during high-stakes trials. Furthermore, the transition condition also exhibited a significant main effect (βtransition = −0.04, 95% CI = [−0.07, −0.01]), signifying that model-based control was less pronounced in variable-transition blocks than in stable-transition blocks.
Figure
2.
The metacontrol effect. Logistic regression weights show the influence of stakes, transitions and their interaction effect on the model-based weights in the two groups. The vertical line represents the 95% confidence interval, and the dots represent the mean.
In contrast, the model-based weights of older adults were not affected by task conditions (βstake = 0.01, 95% CI = [−0.02, 0.04], βtransition = 0.01, 95% CI = [−0.01, 0.04]). These findings align with previous research[6], demonstrating that older adults performed weaker model-based control and exhibited weaker metacontrol than their younger counterparts.
3.2
Older adults are sensitive to stake cues and feedback
To investigate the participants’ sensitivity to rewards, we analyzed their real-time pupil diameter during low- and high-stake trials. Fig. 3a and 3d present a significant difference between the two stake conditions, indicating that high-stake trials elicited larger pupil diameters than low-stake trials. Then, we examined the reaction time under low-stake and high-stake conditions. Consistent with the findings in the pupillary analysis, we observed that high-stake trials led to shorter reaction times than low-stake trials during stage 1 for both younger adults (paired t test: t (28) = 2.69, p =0.01, Fig. 3b) and older adults (paired t test: t (50) = 2.34, p =0.02, Fig. 3c). Additionally, this effect is replicated in stage 2 (younger adults, paired t test: t (28) = 3.1, p =0.004, Fig. 3e; older adults, paired t test: t (50) = 3.07, p =0.003, Fig. 3f).
Figure
3.
Time series depicting the average pupil diameter of low-stake and high-stake trials (a, d) over the course of trials. Solid lines indicate the mean pupil diameter (baseline-corrected). Shaded areas indicate standard errors (SEs) of pupil diameter (baseline-corrected). Red lines on the top indicate time points of a reliable [p <0.05] positive effect for younger and older adults. Reaction time in stage 1 (b, e) and stage 2 (c, f). The error bar represents the standard error of the mean. * indicates p <0.05, ** indicates p <0.01, and *** indicates p <0.001.
These results show that cognitive effort can be adaptively allocated based on stake cues, supporting the notion that older adults exhibit similar sensitivity to reward magnitude as younger adults.
We also analyzed whether participants could differentiate between different types of feedback by observing their pupil responses during the feedback phase. Our findings showed that both groups could distinguish between different feedback sizes based on the changes in pupil diameters (Fig. 4). This finding further supports the notion that older adults are sensitive to rewards.
Figure
4.
Time series depicting the average pupil diameter of high (7–9), medium (3–6) and low (0–2) reward trials in the feedback phase. Solid lines indicate the mean pupil diameter (baseline-corrected). Shaded areas indicate SEs of pupil diameter (baseline-corrected).
3.3
Older adults are not sensitive to structural needs
To investigate whether participants were sensitive to structural needs, we analyzed their real-time pupil diameter during stable-transition and variable-transition blocks. Our findings revealed that in younger adults, the structure needs in the variable-transition block led to an increase in pupil diameter (Fig. 5a and 5d) and in reaction time (stage 1: paired t test: t (28) = −1.8, p =0.08, Fig. 5b; stage 2: paired t test: t (28) = −2.79, p =0.0009, Fig. 5c). However, these changes were not observed in older adults (stage 1, paired t test: t (50) = −0.04, p =0.96, Fig. 5e; stage 2, paired t test: t (50) = 0.26, p =0.79, Fig. 5f).
Figure
5.
Time series depicting the average pupil diameter of stable-transition and variable-transition trials (a, d) over the course of trials. Solid lines indicate the mean pupil diameter. Shaded areas indicate SEs of pupil diameter (baseline-corrected). Red lines on the top indicate time points of a reliable [p < 0.05] positive effect for younger and older adults. Reaction time in stage 1 (b, e) and stage 2 (c, f). The error bar represents the standard error of the mean. *** indicates p <0.001 and NS = not significant.
These results suggest that older adults have a diminished sensitivity to structural needs and renewal, implying that there may be defects in the representation of structural knowledge.
3.4
Subjective report recognizes older adults with metacontrol
Our experimental results suggest that older adults are not sensitive to structure. To confirm that differences in structural knowledge are an essential reason for the decline in metacontrol in older adults, we group participants using subjectively reported difficulty in updating structures. Twenty-eight, 1, and 0 of 29 younger participants reported easy/medium/hard, respectively, compared with 25, 15, and 11 of 51 older participants, respectively. We found significant metacontrol effects under stake conditions (βstake = 0.04, 95% CI = [0.01, 0.07], Fig. 6) among older adults who perceived structural renewal as easy and moderate evidence for metacontrol effects under transition conditions (βtransition = −0.02, 95% CI = [−0.05, 0], with 92% of the posterior mass below zero), which suggests that older adults with good structural knowledge can exert metacontrol similar to younger adults.
Figure
6.
The metacontrol effect after grouping. Logistic regression weights show the influence of stakes, transitions and their interaction effect on the model-based weights in older adults. Older adults were grouped according to subjective reports of different levels of structural update difficulty. The vertical line represents the 95% confidence interval, and the dots represent the mean.
3.5
Metacontrol can be masked by structural knowledge error
To further illustrate the relationship between structural knowledge error and metacontrol, we generated simulated data demonstrating that standard computational models have difficulty detecting metacontrol in participants with inaccurate structural representations. Our approach involves introducing different noise levels to the classical hybrid reinforcement learning model, accompanied by varying degrees of model-based weight adjustment. Subsequently, we use these agents for behavioral simulation and fit the simulated behavior with a model that assumes correct structural knowledge. As anticipated, our simulations reveal that higher levels of structural noise led to less differentiated estimates of model-based weight (Fig. 7). Older adults with inaccurate structural representations may encounter similar challenges.
Figure
7.
Model prediction. Comparison of simulated vs. empirically observed model-based weight differences. Agents were divided into three groups based on structural noise levels. The error bar represents the standard error of the mean.
This study aimed to determine whether cognitive or motivational factors are primarily responsible for weakening metacontrol in older adults. In our metacontrol task, pupillary and behavioral evidence supports that older adults are as sensitive to stake cues and feedback as younger adults, refuting the assumption that older individuals are insensitive to rewards or do not engage in cost‒benefit analysis in sequential decision-making. Then, we found that older adults are minimally affected by structural needs in comparison to younger adults, indicating cognitive deficits in the foundation of model-based decision-making among older individuals. Importantly, older adults with good structural knowledge exercised metacontrol as younger adults, underscoring the role of structural knowledge error in reduced metacontrol in aging. Finally, model simulation shows that if the structure is not expressed correctly, it is difficult for the computational model to detect even if metacontrol is performed.
The theory of the expected value of control[26] describes the computational principles and underlying processes of cognitive control allocation, including monitoring, specification (based on cost‒benefit analysis), and regulation. Based on this theory, metacontrol in decision-making can be decomposed into monitoring the decision-making process, determining the intensity of model-based control to exert, and implementing the specified model-based control. In contrast to previous metacontrol research in aging[6], our study identifies the stages at which errors occur in metacontrol. Specifically, the difference in pupil response and reaction time following different stake cues suggests that cognitive control has been allocated and executed. Therefore, motivational factors may play a secondary role, and the observed reduced metacontrol in older adults can primarily be attributed to the failure of model-based control regulation. This conclusion is consistent with recent findings showing that metacontrol is mainly independent of intrinsic motivation[27].
An explicit structure is essential for model-based decisions[18,28]. However, this cognitive foundation is greatly affected by aging. Consistent with previous research that found that older adults may have age-related deficits in the representation of task transition structures or state spaces[10,11], our study did not find pupillometry evidence that older adults responded to structural needs. These results all indicate that older adults rarely maintain accurate structural knowledge. When the structure is unreliable, it is challenging to regulate model-based control that relies on structure precisely. Moreover, this viewpoint is supported by our model simulation results.
The second explanation for being unable to observe metacontrol in older adults is that they do not use model-based control. For example, previous research has shown that older adults are more prone to using heuristic methods than younger adults[29]. Specifically, an experiment found that older adults use the heuristic “win-stay, lose-shift” strategy instead of the more complex model-based strategy used by younger adults[30]. If older adults adopt heuristic strategies or discard model-based decision-making in sequential decision-making tasks, using the classical dual-system model may lead to a misunderstanding of the relative contributions of different decision strategies and, naturally, a misunderstanding of the changes in relative contributions of the two systems under different contextual demands. Future work should use advanced computational models and matched paradigms to solve this problem.
Our study has several limitations that need to be further optimized in future studies. First, our interpretation of the experimental phenomenon is limited. The current study found that younger adults exhibited earlier and longer pupil dilation after high-stake cues compared to older adults, which may reveal differences between younger and older adults in cognitive effort allocation or preplanning processes[31]. The underlying mechanisms go beyond this study’s scope. Second, although the sample size of our study was comparable to that of previous studies, a larger sample size might better reveal the individual variability of metacontrol.
5.
Conclusions
In summary, our results suggest that metacontrol inefficiency in older adults is primarily due to the inability to learn and maintain structural knowledge rather than reward insensitivity or reduced motivation. The present study sheds light on aging-related impairments in goal-directed behavior[32,33]. We advocate that future studies explore intervention strategies to help older adults flexibly construct task structure representations in daily life activities.
Conflict of Interest
The authors declare that they have no conflict of interest.
A more lightweight and representative scale-in-scale operator, namely SIS is proposed. The operator can be plugged into any promising backbone to replace regular convolution operator.
To be more efficient, an improved SIS series is proposed. The series maintains promising results within nearly half parameters.
Extensive experiments demonstrate the superior performance of our SIS operator series compared with state-of-the-art methods on image classification, object detection, key points estimation and semantic segmentation.
Wang Q, Chen W, Wu X, et al. Detail preserving multi-scale exposure fusion. In: 2018 25th IEEE International Conference on Image Processing (ICIP). Athens, Greece: IEEE, 2018: 1713-1717.
[2]
Wang B, Lei Y, Li N, et al. Multi-scale convolutional attention network for predicting remaining useful life of machinery. IEEE Transactions on Industrial Electronics,2021, 68 (8): 7496–7504. DOI: 10.1109/TIE.2020.3003649
[3]
Yu J, Xie H, Xie G, et al. Multi-scale densely U-Nets refine network for face alignment. In: 2019 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). Shanghai, China: IEEE, 2019: 691–694.
[4]
Zhang X, Zhang W. Application of new multi-scale edge fusion algorithm in structural edge extraction of aluminum foam. IEEE Access,2020, 8: 15502–15517. DOI: 10.1109/ACCESS.2019.2963454
[5]
Liu G, Wang C, Hu Y. RPN with the attention-based multi-scale method and the adaptive non-maximum suppression for billboard detection. In: 4th International Conference on Computer and Communications. Chengdu, China: IEEE, 2018: 2018.8780907.
[6]
Sun K, Xiao B, Liu D, et al. Deep high-resolution representation learning for human pose estimation. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA: IEEE, 2019: 5686–5696.
[7]
Fang X, Yan P. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging,2020, 39 (11): 3619–3629. DOI: 10.1109/TMI.2020.3001036
[8]
Huang G, Chen D, Li T, et al. Multi-scale dense networks for resource efficient image classification. In: 6th International Conference on Learning Representations. Vancouver, Canada: ICLR, 2018.
[9]
Moukari M, Picard S, Simon L, et al. Deep multi-scale architectures for monocular depth estimation. In: 2018 25th IEEE International Conference on Image Processing. Athens, Greece: IEEE, 2018: 2940–2944.
[10]
Papyan V, Elad M. Multi-scale patch-based image restoration. IEEE Transactions on Image Processing,2016, 25 (1): 249–261. DOI: 10.1109/TIP.2015.2499698
[11]
Li J, Fang F, Li J, et al. MDCN: Multi-scale dense cross network for image super-resolution. IEEE Transactions on Circuits and Systems for Video Technology,2021, 31 (7): 2547–2561. DOI: 10.1109/TCSVT.2020.3027732
[12]
Lowe D G. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision,2004, 60 (2): 91–110. DOI: 10.1023/B:VISI.0000029664.99615.94
[13]
Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA: IEEE, 2015: 1–9.
[14]
Gao S, Cheng M. M, Zhao K, et al. Res2Net: A new multi-scale backbone architecture. IEEE Transactions on Pattern Analysis and Machine Intelligence,2019, 43 (2): 652–662. DOI: 10.1109/TPAMI.2019.2938758
[15]
Krizhevsky A, Sutskever I, Hinton G. E. ImageNet classification with deep convolutional neural networks. In: Proceedings of the 25th International Conference on Neural Information Processing Systems: Volume 1. Red Hook, NY: Curran Associates Inc, 2012: 1097–1105.
[16]
Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. https://arxiv.org/abs/1409.1556.
[17]
He K, Zhang X, Ren S, et al. Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV: IEEE, 2016: 770-778.
[18]
Xie S, Girshick R, Dollár P, et al. Aggregated residual transformations for deep neural networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, 2017: 5987–5995.
[19]
Huang G, Liu Z, Van Der Maaten L, et al. Densely connected convolutional networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, 2017: 2261–2269.
[20]
Iandola F N, Han S, Moskewicz M W, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5 MB model size. In: 5th International Conference on Learning Representations. Toulon, France: ICLR, 2017.
[21]
Howard A G, Zhu M, Chen B, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications. https://arxiv.org/abs/1704.04861.
[22]
Zhang X, Zhou X, Lin M, et al. ShuffleNet: An extremely efficient convolutional neural network for mobile devices. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, 2018: 6848–6856.
[23]
Lin M, Chen Q, Yan S. Network in network. In: ICLR 2014 Conference.Banff, Canada: ICLR, 2014.
[24]
Sun S, Pang J, Shi J, et al. FishNet: A versatile backbone for image, region, and pixel level prediction. Advances in Neural Information Processing Systems 31 (NeurIPS 2018). Montréal, Canada: NeurIPS, 2018: 754–764.
[25]
Yu F, Wang D, Shelhamer E, et al. Deep layer aggregation. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, 2018: 2403–2412.
[26]
Liao R, Zhao Z, Urtasun R, et al. LanczosNet: Multi-scale deep graph convolutional networks. In: Seventh International Conference on Learning Representations. New Orleans, LA: ICLR, 2019.
[27]
Liu Y, Wang Y, Wang S, et al. CBNet: A novel composite backbone network architecture for object detection. Proceedings of the AAAI Conference on Artificial Intelligence,2020, 34 (7): 11653–11660. DOI: 10.1609/aaai.v34i07.6834
[28]
Chen L, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Computer Vision–ECCV 2018. Cham, Switzerland: Springer, 2018: 833-851.
[29]
Xiao B, Wu H, Wei Y. Simple baselines for human pose estimation and tracking. In: Computer Vision–ECCV 2018. Cham, Switzerland: Springer, 2018: 472–487
Figure
1.
Comparison between bottleneck modules(a), Res2Net module(b) and SIS module series(c, d, e, f, g) and illustration of scale-in-scale mechanism (h).
Figure
2.
Visual comparisons between Faster RCNN and SIS equipped Faster RCNN.
References
[1]
Wang Q, Chen W, Wu X, et al. Detail preserving multi-scale exposure fusion. In: 2018 25th IEEE International Conference on Image Processing (ICIP). Athens, Greece: IEEE, 2018: 1713-1717.
[2]
Wang B, Lei Y, Li N, et al. Multi-scale convolutional attention network for predicting remaining useful life of machinery. IEEE Transactions on Industrial Electronics,2021, 68 (8): 7496–7504. DOI: 10.1109/TIE.2020.3003649
[3]
Yu J, Xie H, Xie G, et al. Multi-scale densely U-Nets refine network for face alignment. In: 2019 IEEE International Conference on Multimedia & Expo Workshops (ICMEW). Shanghai, China: IEEE, 2019: 691–694.
[4]
Zhang X, Zhang W. Application of new multi-scale edge fusion algorithm in structural edge extraction of aluminum foam. IEEE Access,2020, 8: 15502–15517. DOI: 10.1109/ACCESS.2019.2963454
[5]
Liu G, Wang C, Hu Y. RPN with the attention-based multi-scale method and the adaptive non-maximum suppression for billboard detection. In: 4th International Conference on Computer and Communications. Chengdu, China: IEEE, 2018: 2018.8780907.
[6]
Sun K, Xiao B, Liu D, et al. Deep high-resolution representation learning for human pose estimation. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Long Beach, CA: IEEE, 2019: 5686–5696.
[7]
Fang X, Yan P. Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction. IEEE Transactions on Medical Imaging,2020, 39 (11): 3619–3629. DOI: 10.1109/TMI.2020.3001036
[8]
Huang G, Chen D, Li T, et al. Multi-scale dense networks for resource efficient image classification. In: 6th International Conference on Learning Representations. Vancouver, Canada: ICLR, 2018.
[9]
Moukari M, Picard S, Simon L, et al. Deep multi-scale architectures for monocular depth estimation. In: 2018 25th IEEE International Conference on Image Processing. Athens, Greece: IEEE, 2018: 2940–2944.
[10]
Papyan V, Elad M. Multi-scale patch-based image restoration. IEEE Transactions on Image Processing,2016, 25 (1): 249–261. DOI: 10.1109/TIP.2015.2499698
[11]
Li J, Fang F, Li J, et al. MDCN: Multi-scale dense cross network for image super-resolution. IEEE Transactions on Circuits and Systems for Video Technology,2021, 31 (7): 2547–2561. DOI: 10.1109/TCSVT.2020.3027732
[12]
Lowe D G. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision,2004, 60 (2): 91–110. DOI: 10.1023/B:VISI.0000029664.99615.94
[13]
Szegedy C, Liu W, Jia Y, et al. Going deeper with convolutions. In: 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA: IEEE, 2015: 1–9.
[14]
Gao S, Cheng M. M, Zhao K, et al. Res2Net: A new multi-scale backbone architecture. IEEE Transactions on Pattern Analysis and Machine Intelligence,2019, 43 (2): 652–662. DOI: 10.1109/TPAMI.2019.2938758
[15]
Krizhevsky A, Sutskever I, Hinton G. E. ImageNet classification with deep convolutional neural networks. In: Proceedings of the 25th International Conference on Neural Information Processing Systems: Volume 1. Red Hook, NY: Curran Associates Inc, 2012: 1097–1105.
[16]
Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. https://arxiv.org/abs/1409.1556.
[17]
He K, Zhang X, Ren S, et al. Deep residual learning for image recognition. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV: IEEE, 2016: 770-778.
[18]
Xie S, Girshick R, Dollár P, et al. Aggregated residual transformations for deep neural networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, 2017: 5987–5995.
[19]
Huang G, Liu Z, Van Der Maaten L, et al. Densely connected convolutional networks. In: 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, HI: IEEE, 2017: 2261–2269.
[20]
Iandola F N, Han S, Moskewicz M W, et al. SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5 MB model size. In: 5th International Conference on Learning Representations. Toulon, France: ICLR, 2017.
[21]
Howard A G, Zhu M, Chen B, et al. Mobilenets: Efficient convolutional neural networks for mobile vision applications. https://arxiv.org/abs/1704.04861.
[22]
Zhang X, Zhou X, Lin M, et al. ShuffleNet: An extremely efficient convolutional neural network for mobile devices. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, 2018: 6848–6856.
[23]
Lin M, Chen Q, Yan S. Network in network. In: ICLR 2014 Conference.Banff, Canada: ICLR, 2014.
[24]
Sun S, Pang J, Shi J, et al. FishNet: A versatile backbone for image, region, and pixel level prediction. Advances in Neural Information Processing Systems 31 (NeurIPS 2018). Montréal, Canada: NeurIPS, 2018: 754–764.
[25]
Yu F, Wang D, Shelhamer E, et al. Deep layer aggregation. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, 2018: 2403–2412.
[26]
Liao R, Zhao Z, Urtasun R, et al. LanczosNet: Multi-scale deep graph convolutional networks. In: Seventh International Conference on Learning Representations. New Orleans, LA: ICLR, 2019.
[27]
Liu Y, Wang Y, Wang S, et al. CBNet: A novel composite backbone network architecture for object detection. Proceedings of the AAAI Conference on Artificial Intelligence,2020, 34 (7): 11653–11660. DOI: 10.1609/aaai.v34i07.6834
[28]
Chen L, Zhu Y, Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Computer Vision–ECCV 2018. Cham, Switzerland: Springer, 2018: 833-851.
[29]
Xiao B, Wu H, Wei Y. Simple baselines for human pose estimation and tracking. In: Computer Vision–ECCV 2018. Cham, Switzerland: Springer, 2018: 472–487
Open Access
DownLoad:
