Investigated neural network (NN) structures with corresponding weight and bias count, size to store the network and achieved accuracy. The quantized two-layer NN was used on the IoT nodes. The maximum difference in accuracy between quantized NN and its full precision correspondence was 0.07% and was observed in both directions, better and worse than the full precision network.

	Layers	Weights and biases	Computations per inference	Accuracy [%]
Deep six-layer NN by Cireşan et al.64	784–2500–2000–1500–1000–500–10	∼12 million (∼46 MB)	∼24 million	99.65
Large two-layer NN (15 epochs)	784–800–10	636 010 (∼2.5 MB)	1 270 400	98.3
Small two-layer NN (5 epochs)	784–64–10	50 890 (∼200 kB)	101 632	97
Two-layer NN on small images (5 epochs)	196–32–10	6634 (∼26 kB)	13 184	95.00 ± 0.17
Quantized two-layer NN on small images	196–32–10	6634 (∼6.5 kB)	13 184	94.99 ± 0.16