Table 2. Hyper-parameters for neural network architecture search.

To form candidate networks, first a number of layers (per type) is chosen, ranging from 2 to 8 (in multiples of 2) for spatial-temporal models and 1–4 for the spatiotemporal and long short-term memory (LSTM) ones. Next, a spatial and temporal kernel size per Layer is picked where relevant, which remains unchanged throughout the network. For the spatiotemporal model, the kernel size is equal in both the spatial and temporal directions in each layer. Then, for each layer, an associated number of kernels/feature maps is chosen such that it never decreases along the hierarchy. Finally, a spatial and temporal stride is chosen. For the LSTM networks, the number of recurrent units is also chosen. All parameters are randomized independently and 50 models are sampled per network type. Columns 2–4: Hyper-parameter values for the top-performing models in the ART. The values given under the spatial rows count for both the spatial and temporal directions for the spatiotemporal model.

	Hyper-parameters	Spatial-temporal	Spatiotemporal	LSTM
Num. layers	[1, 2, 3, 4]	4+4	4	3+1
Spatial kernels (pL)	[8, 16, 32, 64]	[8,16,16,32]	[8, 8, 32, 64]	[8, 16, 16]
Temporal kernels (pL)	[8, 16, 32, 64]	[32, 32, 64, 64]	n/a	n/a
Spatial kernel size	[3, 5, 7, 9]	7	7	3
Temporal kernel size	[3, 5, 7, 9]	2	n/a	n/a
Spatial stride	[1, 2]	9	2	1
Temporal stride	[1, 2, 3]	3	n/a	n/a
Num. recurrent units	[128, 256]	n/a	n/a	256