1. State of the art
1.1. Neuroscience to ML
1.1.1. No clue what methods are appropriate to ML (Machine learning papers naive way)
1.1.1.1. Capsulles
1.1.1.2. HTM
1.1.2. biological equivalences of backpropagation
1.1.3. close spatial and temporal section correlation
1.1.4. neuroevolution as a mechanism
1.1.5. Neural Turing Machines
1.2. Raise of DL and RF
1.2.1. Spiking Neural Networks
1.2.2. NC
1.2.2.1. very-large-scale integration (VLSI)
1.2.2.2. spiking neuron
1.2.2.2.1. comparison CNN SNN
1.2.2.3. HTM
1.3. Towards embedded
1.3.1. sensor-fusion and on-line processing
1.3.1.1. Examples of AlexNet GMACS...
1.3.1.2. RESNet
1.3.2. Network models compression
1.3.2.1. NO EXPLANATIONS - extract from papers and show limitations (they lack synergies - they reach their limits, it is clear how to compress it but do not know how to combine with architecture modification = reduce precioison but do not know what further, no automatic procedure)
1.3.2.2. optimization of network architecture
1.3.2.3. optimization of the problem
1.3.2.4. minimization of the number of bits
1.3.2.5. optimization of basic neuron and network hardware
1.3.2.5.1. capsulles Hinton
1.3.2.6. Networks compresion techniques examples
2. NEW STATE
2.1. a thorough review of what has been done
2.1.1. Machine learning
2.1.1.1. missing points from Neuroscience (what to take, what to corporate)
2.1.1.1.1. INSCApe (gradually systematic way). methodoloy with architecture
2.1.1.2. very short intro about developmnts in nn in general (cnns, rnns)
2.1.1.3. training
2.1.1.4. precision reduction
2.1.1.5. network compression
2.1.1.6. architecture considerations (inception, resnet etc.)
2.1.2. Neuroscience
2.1.2.1. missing points to ML
2.1.3. Neuromorphing
2.2. explanantion why and where the field is striving to
2.2.1. towards GAI
2.3. Limitation and challenges
2.3.1. Move to hardware