(A) Top, a typical ’raw’ extracellular recording from a single CA1 electrode. Bottom, wavelet decomposition of the same data, power shown for frequency bands from 2 Hz to 15 kHz (bottom to top row). (B) At each timestep wavelet coefficients (64 time points, 26 frequency bands, 128 channels) were fed to a deep network consisting of 2D convolutional layers with shared weights, followed by a fully connected layer with a regression head to decode self-location; schematic of architecture shown. (C) Example trajectory from R2478, true position (black) and decoded position (blue) shown for 3 s of data. Full test-set shown in Figure 1—video 1. (D) Distribution of decoding errors from trial shown in (C), mean error (14.2 cm ± 12.9 cm, black), chance decoding of self-location from shuffled data (62.2 cm ± 9.09 cm, red). (E) Across all five rats, the network (CNN) was more accurate than a machine learning baseline (SVM) and a Bayesian decoder (Bayesian) trained on action potentials. This was also true when the network was limited to high-frequency components (>250Hz, CNN-Spikes). When only local frequencies were used (<250Hz, CNN-LFP), network performance dropped to the level of the Bayesian decoder (distributions show the fivefold cross validated performance across each of five animals, n=25). Note that this likely reflects excitatory spikes being picked up at frequencies between 150 and 250 Hz (Figure 1—figure supplement 2). (F) Decoding accuracy for individual animals, the network outperformed the Bayesian decoder in all cases. An overview of the performance of all tested models can be seen in Figure 1—figure supplement 3. (G) The advantage of the network over the Bayesian decoder increased when the available data was reduced by downsampling the number of channels (data from R2478). Inset shows the difference between the two methods.