ML - The way the world works - analyzing how things work

Thought vector

Weekly thoughts

Thoughts

User experience writing code with the machine

Forget, memory, hide or transfer cell gates of lstm

spiking networks the brain has very low power density and very low frequency. deep learning uses in-memory computing. spike-timing dependent plasticity timing between the spikes is how learning is taking place in the synapse. The synapse is connected to the pre-synaptic neuron and a post-synaptic neuron. The Input spikes from the pre-synaptic neuron and the post synaptic neuron affect the weights on the synapse. The delay between the spikes determines who the synapse learns. The synapse will depress from negative current with voltage flowing from pre neuron to the post neuron. if the pre to post voltage is positive you have excitation of the synapse. feedforward network creates a perceptron. multiple synapse connect to a post neuron. the perceptron is self learning. the artificial neurons learn like the biological neurons. the artificial neuron can memorize objects. recurrent networks represent feedback systems. in the recurrent network all neurons talk with each other. the recurrent network has an inhibitor synapse and an excitor synapse. hebbian learning occurs: neurons that fire together wire together.  spatiotemporal networks allow for sequence learning and recognition. spatiotemporal networks learn sequence over time. potentiation is "the increase in strength of nerve impulses along pathways which have been used previously, either short-term or long-term".

Quantum scientist have also shown that an array of Single electron Transistor - SETs create a form of neural network." SETs construct computers that use individual electrons to carry information. SET biggest problem is operating at room temperature. Quantum tunneling means the can "interact capacitively rather than by current flow throught the wires." "When their interactions result from the quantum tunneling of electrons, quantum dots can collectively behave as a form of quantum cellular automaton, QCA. QCA computers may show associative memory. If Decoherence can be avoid a qbit can form with a 0 or 1 or superposition state of both at the same time. 5 qbits could handle 32 states (2^n), simulateously; a conventional computer would handle 32 sets of 5 bits, or 160 bits in all. 64 bit encryption could be processed with one 64 qubit operation, whereas, a conventional computer requiring 2^64, 1.84 x10^19 operations or 292.5 years, 18 billion billion times more powerful than a 64 bit binary compute

In Adiabatic QC, you evolve the system under a time-dependent Hamiltonian. Two constraints are your initial ground state and the Hamiltonian changes very slowly in time, so for all times your quantum state remains close to the ground

bay labs focuses on bringing ai to healthcare by studying medical imaging using deep learning. baylabs wants to increase the quality, value and access to medical imaging. medical image is used to detect health defects. baylabs asks how can we bring more imaging to more people. The us spends $100 billion annually on medical image. baylabs focuses on ultrasound for medical imaging. ultrasound is very affordable and imaging can be very good. Today, ultrasound devices are a hand held probe and a tablet for display.  What does it mean to intrepret ultrasound images? experts recognize certain biological structures and certain defect features. They then build a story of what is going on in the image. In 2013, machines began recognizing object at a glance better than humans it neurons. if neural networks are so good at recognizing object outside the body then apply them to objects within the body. The baylabs object recognition can be put on the same device gather ultrasound images. bay labs used technology built in hardware by philips lumen to put their software. Bay labs distributes its software to medical sites around the worlds as low cost medical imaging.

Thoughts