Extended Data Fig. 3: Decoding neuronal representations of the first-odour identity and choice across different task epochs.
From: Volatile working memory representations crystallize with practice
(a) Visualization of the first odour identity and choice during different task epochs using t-distributed stochastic neighbour embedding (t-SNE), in an expert animal performing at 92%. The decoding accuracy of the SVM decoder, presented in the right panel, illustrates the presence of working memory information in the neuronal population activity during the late-delay epoch for the same dataset. (b) Decoding accuracy for odour identification in the go/no-go task, devoid of a working memory component (n = 3 mice). (c) Distribution of the animal’s licking behaviour during the go/no-go task (n = 3 mice). (d) Left panel: Decoding accuracy for the first odour in normal and confusion tasks with randomized reward contingencies, using N% of the top selective neurons during first-odour and late-delay epochs. Notably, in the confusion task, decoding accuracy during the late-delay epoch approaches chance level (n = 3 mice). Right panel: Evaluation of choice decoding in both normal and confusion tasks using the activity of all neurons. Shuffled data pertains to the confusion task. (e) Behavioural performance of mice at expert proficiency levels, featuring the identical mice analysed in Fig. 3g-j (p = 0.25; one-way ANOVA). (f) Number of imaged neurons for each imaging day, showcasing the same mice examined in Fig. 3g-j (p = 0.25; one-way ANOVA). (g) Percentage of neurons exhibiting selectivity in each time epoch on every imaging day, featuring the identical mice analysed in Fig. 3g-j (first-odour epoch p = 0.69, late-delay epoch p = 0.46, choice epoch p = 0.34; one-way ANOVA). (h) Decoding accuracy for the first odour using N% of choice epoch’s most selective neurons. (i) Decoding accuracy in predicting the initial odour identity and choice based on the separate and combined activity of M1 and M2 neurons using a non-linear Long Short-term Memory (LSTM) recurrent neural network decoder. Equal numbers of M1 and M2 neurons are employed in the decoding process to ensure a fair comparison. (j) Similar to (i), but using simultaneous recordings of the retrosplenial and secondary motor cortices. (k) Decoding accuracy for the first odour (left two) or choice (right) using N% of selective neurons during the first-odour (left), late-delay (middle), or choice (right) epochs. The analysis is performed with a non-linear LSTM decoder using the identical data from n = 5 mice performing at expert level showcased in Fig. 2e. (l) Maximum decoding accuracy during the first-odour, late-delay, or choice epochs using N% of each epoch selective neurons. Dashed lines represent decoding from shuffled data. (m) Trial category prediction accuracy (AC, AD, BC, BD; n = 5 mice). Chance level, represented by the red trace, is 25% due to the four possible categories. (a)–(m), Mean ± s.e.m.
