Evolution of neural activity in circuits bridging sensory and abstract knowledge

  1. Francesca Mastrogiuseppe  Is a corresponding author
  2. Naoki Hiratani
  3. Peter Latham
  1. Gatsby Computational Neuroscience Unit, University College London, United Kingdom
  2. Center for Brain Science, Harvard University, United States
8 figures, 2 tables and 1 additional file

Figures

Schematics of tasks and circuit model used in the study.

(A) Illustration of the two categorization tasks. In the simple categorization task, half the stimuli are associated with category A and the other half with category B. In the context-dependent …

Figure 2 with 4 supplements
Characterization of activity evolution during the simple categorization task.

Results from simulations. The first column (A–D) shows a naive circuit (pre-learning); the second (E–H) and third (I–L) columns show two trained circuits (post-learning), characterized by different …

Figure 2—figure supplement 1
Characterization of activity evolution during the simple categorization task; additional results, part I.

(A) Extensive results from simulations. Each point represents a model. We simulated 5000 different models, with parameters drawn randomly and uniformly from the following ranges: the number of …

Figure 2—figure supplement 2
Characterization of activity evolution during the simple categorization task; additional results, part II.

(A–D) Population response to categories A and B averaged over stimuli. Each dot represents a neuron. Panels A, B and C, D display two sample circuits, different from those displayed in Figure 2. …

Figure 2—figure supplement 3
Learning curves.

Behaviour of the loss function (Equation 9) over learning epochs in four sample networks. (A, B) refer to the simple categorization task; parameters are, respectively, as in Figure 2E–H and Figure …

Figure 2—figure supplement 4
Simple categorization task with structured inputs and heterogeneity.

Analysis of category correlation for a naive circuit (A), and two trained ones (B, C). Details are, respectively, as in Figure 2C–D, G–H and K–L. (D) Extensive analysis of average category …

Figure 3 with 2 supplements
Analysis of activity evolution during the simple categorization task.

Results from mathematical analysis. (A–C) Cartoons illustrating how activity evolves over learning. The three columns are as in Figure 2: pre-learning (first column) and post-learning for two …

Figure 3—figure supplement 1
Comparison between finite-size networks and approximate mathematical description for the simple categorization task, part I.

Dashed lines show the theoretical predictions; dots show the average over 400 simulations where both the initial connectivity and the sensory inputs were drawn at random. Error bars show the …

Figure 3—figure supplement 2
Comparison between finite-size networks and approximate mathematical description for the simple categorization task, part II.

Details in (A-C) are as in Figure 3—figure supplement 1. We used different parameters (see Table 2), which lead to positive category correlation.

Category correlation depends on circuit and task properties.

(A) Category correlation as a function of the threshold and gain of the readout neuron. Grey arrows indicate the threshold and gain that are used in panels C and D. The learning rate ratio, ηw/ηu, is …

Magnitude of category selectivity depends on connectivity with the readout neuron.

(A–C) Category selectivity as a function of the initial readout connectivity w0,i (in absolute value). The three columns are as in Figure 2: pre-learning (first column) and post-learning for two …

Figure 6 with 3 supplements
Characterization of activity evolution during the context-dependent categorization task.

Results from simulations. The first column (A–C) shows a naive circuit (pre-learning); the second (D–F) and third (G–I) columns show two trained circuits (post-learning), characterized by different …

Figure 6—figure supplement 1
Characterization of activity evolution during the context-dependent categorization task; additional results, part I.

(A) Extensive results from simulations. Each point represents a model. We simulated 5000 different models, with parameters drawn randomly. Details as in Figure 2—figure supplement 1A, except that …

Figure 6—figure supplement 2
Characterization of activity evolution during the context-dependent categorization task; additional results, part II.

(A, B) Population response to categories A and B, averaged over trials. Each dot represents a neuron. Panels A and B correspond to two sample circuits, characterized by different values of the …

Figure 6—figure supplement 3
Characterization of activity evolution during the context-dependent categorization task; additional results, part III.

(A-B) Changes in category (panel A) and context (panel B) selectivity as a function of the components on the category and context directions, dicat and dictx (in absolute value). The latter are defined in …

Figure 7 with 1 supplement
Analysis of activity evolution during the context-dependent categorization task.

Results from mathematical analysis. (A–C) Cartoons illustrating how activity evolves over learning. Orange and blue symbols are associated with categories A and B, respectively; circles and squares …

Figure 7—figure supplement 1
Comparison between finite-size networks and approximate mathematical description for the context-dependent categorization task.

Details are as in Figure 3—figure supplement 1. In all panels, the top and bottom plots show, respectively, results for the synaptic drive k and the activity y. (A) Average category selectivity. …

Patterns of pure and mixed selectivity to category and context.

(A) Changes in category selectivity (left) and context selectivity (right) as a function of the initial readout connectivity, w0,i (in absolute value). Details as in Figure 5B, C. (B) Changes in …

Tables

Table 1
Table of parameters for figures in the main text.
FigureNQηw/ηuΘ1,ΨΘ2,ΨΘ1,ΦΘ2,Φ
Figures 2, 3 and 5, first and second columns200200.01.02.01.00.0
Figures 2, 3 and 5, third column200200.01.02.01.02.0
Figure 4A200200.42.02.0variesvaries
Figure 4B200200.4variesvaries1.02.0
Figure 4C20020varies2.02.01.0varies
Figure 4D200varies0.42.02.01.0varies
Figure 6A–C, first and second columns6008 (P = 64)0.01.00.01.00.0
Figure 6A–C, third column6008 (P=64)0.01.00.01.04.0
Figure 7D6008 (P = 64)0.22.52.0variesvaries
Figure 7E6008 (P = 64)varies2.52.01.0varies
Figure 7F600varies0.22.52.01.0varies
Table 2
Table of parameters for figure supplements.
Figure supplementNQηw/ηuΘ1,ΨΘ2,ΨΘ1,ΦΘ2,Φ
Figure 2—figure supplement 1A200variesvariesvariesvariesvariesvaries
Figure 2—figure supplement 2E200200.01.02.01.00.0
Figure 3—figure supplement 1, first columnvaries200.01.00.01.00.0
Figure 3—figure supplement 1, second column20020varies1.00.01.00.0
Figure 3—figure supplement 1, third column200200.01.0varies1.00.0
Figure 3—figure supplement 2, first columnvaries200.01.00.01.02.0
Figure 3—figure supplement 2, second column20020varies1.00.01.02.0
Figure 3—figure supplement 2, third column200200.01.0varies1.02.0
Figure 2—figure supplement 4A, B200120.01.02.01.00.0
Figure 2—figure supplement 4C200120.01.02.01.02.0
Figure 2—figure supplement 4D, E, first column200120.12.0varies1.0varies
Figure 2—figure supplement 4D, E, second column20012varies2.02.01.0varies
Figure 2—figure supplement 4D, E, third column200varies0.12.02.01.0varies
Figure 6—figure supplement 1A, B600variesvariesvariesvariesvariesvaries
Figure 6—figure supplement 2A6008 (P = 64)0.01.03.01.00.0
Figure 6—figure supplement 2B6008 (P = 64)0.01.03.01.04.0
Figure 6—figure supplement 2C600variesvariesvariesvariesvariesvaries
Figure 7—figure supplement 1A–C600varies0.01.00.01.00.0
Figure 7—figure supplement 1D–F600varies0.01.00.01.04.0

Additional files

Download links