Fig 1.
Effect of domain expertise on memorising positions in chess.
A: Top, Chess board configurations from real game settings (game configurations). Bottom, reconstructions of the configurations from memory. These configurations are individual samples generated by the expert model based on the encoding of the presented configurations. Green frames indicate correctly reconstructed pieces, red frames indicate positions where a piece is missing or erroneously appears in the reconstructed game; purple frames indicate pieces whose identity is switched in the reconstruction. B: Same as A but instead of game configurations randomly shuffled pieces are presented (random configurations) and reconstructed. C: Reconstruction accuracy of the model for game and random configurations as a function of the training size. D: Reconstruction accuracy of human participants as a function of chess skill. Data reproduced from [37].
Fig 2.
Memory distortions for lists of words.
A: Frequency of recall from 100 samples for individual words in the text model trained on Wikipedia for the word list associated with ‘music’ from the DRM paper [39]. The lure word (red) is characterised by a recall probability comparable to the studied words (green). B: Comparison of recall probabilities for studied and lure words in the text-VAE model for 10 word lists (left, see Methods), and experiment (right) Roediger et al. [39].
Fig 3.
Context effects on reconstruction of line drawing from memory.
A: Middle column, Ambiguous line drawings from the QuickDraw data set of eyeglasses and dumbbells. Left and right columns, Reconstructions of the image from memory in the dumbbell and eyeglasses contexts, respectively. Context is modelled by using a sketch-VAE trained on sketches from a single category with beta = 2. B: Examples of ambiguous drawings (middle column) and their reconstructions (side columns) when cues are provided to participants (shown as text labels). Data is reproduced from [6]. C: Effect of contextual information on the visual features in recalled stimuli in the model. Quantitative changes (top), qualitative changes (middle), and subtle changes in characteristics (bottom) occur as a result of contextual recall. D: Quantitative changes in visual features with changing context (proportion of the length of the line connecting circular features in the eyeglasses and dumbbell contexts) in the Sketch-RNN model (left) and experiment (right). Experimental data reproduced from [42].
Fig 4.
Rate distortion trade-off in memory for sketch drawings.
A: Illustration of stimulus reconstructions as changes in β result in different points on the rate distortion curve for the sketch-VAE model. Inset image is used as input and is reconstructed with various levels of compression. Optimal forgetting implies moving along the curve in the direction of increasing β corresponding to increasingly prototypical reconstructions of the original drawing. B: Proportion of recalled sketches judged to show category specific distortions in humans due to the context presented during learning at different delays between stimulus presentation and recall. Distortions were evaluated by two of the experimenters and one judge naive to the purpose of the experiment. Data reproduced from [42]. C: Proportion of sketches reconstructed by the model showing category specific distortion as a function of increasing compression. Quantitative changes in visual features are assessed, similar to Fig 3D.
Fig 5.
Rate distortion trade-off in forgetting over time.
A,B: Using the text-VAE model, we modelled the dependence of memory representations on time by gradually increasing the compression rate, β. Recall probabilities averaged over multiple word lists of studied and lure words was measured on a text-VAE trained on Wikipedia entries (top) and on a synthetic vocabulary (bottom). Increasing the compression rate results in a monotonically decreasing recall performance for studied words. In contrast, increased delay of recall leads to an increase in false memories. For critical NS (lure) words the recall probability initially increases with larger compression rates but very high compression rates result in losing gist-like recall as well. Asymptotically the performance on semantically related S words will approach the performance on random word lists as less and less of the structure of the data is used. C,D: Difference between recall probabilities for lure words and studied words as a function of the delay between recall and study for the model (top) and experiments (bottom). Both Wikipedia-trained and synthetic vocabulary trained models predict persistence of false recall of non-studied lure words as compared to studied words, visible as an increase in the difference between lure and studied word recall rates as a function of time. For even longer delays, the gist information is progressively forgotten as well and consequently recall rates for both lure and studied words approach zero. The same pattern of increasing rate up to a delay of three weeks and a subsequent decrease can be observed in experimental data. Data is reproduced from Toglia et al. [45], Seamon et al. [46] and Thapar & McDermott [47].