AI 2041 Flashcards
NLP
NLP: Natural Language Processing
NLP is sub branch of artificial intelligence; natural language refers to the language of humans (speech/writing/nonverbal communication)
Turing Test
test of machine intelligence that hinges on whether an NLP software is capable of fooling humans into thinking that it, too, is human.
GAN’s (Generative Adversarial Networks)
Deepfakes are built on a technology called GAN’s. GAN’s is a pair of adversarial deep learning neural networks.
The first network, the forger network, tried to generate something looks real (ex: a dog). The other network, the detective network, compares the forger’s synthesized dog picture with genuine dog picture to determine if the dog is real or fake. Based on the Detective Network’s feedback, the forger network retrains itself with the goal of fooling the Detective network the next time around. The forger network addresses itself to minimize “loss function” or the difference between its generated images and real images. Then the detective network retrains itself to make the forgeries detectable by maximizing the loss function. These two processes repeat for up to millions of times, with forger and detective both improving their skills until an equilibrium is reached.
CNN: Convolutional Neural Networks for computer vision
Convolutional neural networks at its lowest level is a large number of filters, which are applied repeatedly across an image. Each of the filters can only see only small contiguous sections of the image, just like the receptive fields. Deep learning, through optimizing across many images, decides what each filter learns. Each filter will output its confidence that it saw the particular feature (like a black line) that the filter represents. A CNN’s the higher levels are hierarchially organized, like the neo cortex. The higher levels will take the confidence outputs from the lowerlevels and detect more complex features. For example, if an image of a zebra is fed into a CNN, then the lower level filters might look for the black and white lines in each region of the image. And the higher levels might see stripes and ears and legs in the larger regions. At the highest level, parts of the CNN might specifically try to distinguish zebras horses or tigers.
RPA (robotic process automation)
quantum computer
A quantum computer is a new computer architecture that uses quantum mechanics to perform certain tasks of computation much more efficiently than a classical computer can. Classical computers are based on “bits” (an on/off switch of 0 or 1). A bit it’s like a switch it can either be zero or one. Every app website or for photograph is made out of of of of these bits using binary bits makes classical computers easier to vote and control but also limit their potential for taking on really hard computer programs. Instead of bits, QCs use quantum bits or qbits which are typically subatomic particles such as electrons or photons
superposition
Qbits follow principles of quantum mechanics regarding how atomic and subatomic particles behave, when which include unusual properties that get them super processing capabilities. The first such property is superposition, or the capability for each Qbit to be in multiple states at any given time. This allows multiple cubits and superstition to process a fast number of outcomes
Quantum superposition is a fundamental principle of quantum mechanics. It states that, much like waves in classical physics, any two (or more) quantum states can be added together (“superposed”) and the result will be another valid quantum state; and conversely, that every quantum state can be represented as a sum of two or more other distinct states. Mathematically, it refers to a property of solutions to the Schrödinger equation; since the Schrödinger equation is linear, any linear combination of solutions will also be a solution.
Entanglement
The second property is entanglement, which means to cubits remain connected so that the actions performed on one affect the other, even when separate by great distances. Thanks for entanglements, every cubit added to a quantum can machine exponentially increases its computing power. 2 double of 100,000,000 classical super computer, you would have to spend another hundred million dollars. To double your quantum computing, you just need to add one more cubit.
Quantum entanglement is a physical phenomenon that occurs when a group of particles are generated, interact, or share spatial proximity in a way such that the quantum state of each particle of the group cannot be described independently of the state of the others, including when the particles are separated by a large distance. The topic of quantum entanglement is at the heart of the disparity between classical and quantum physics: entanglement is a primary feature of quantum mechanics lacking in classical mechanics.
Sequence Transduction
SELF-SUPERVISED GENERAL
NLP
Recently, however, a simple but elegant new approach for self-supervised learning emerged. Self-supervised learning means Al supervises itself, and no human labeling is required, thus overcoming bottleneck.
This approach is called “sequence transduction.” To train
a sequence-transduction neural network, the input is simply the se-
quence of all the words up to a point, and the output is simply the se-
quence of words after that point. For example, an input of “four score and
seven years ago”
‘ will transduce the predictive output “our fathers brought
forth, upon this continent.” You probably already use simple versions of this every day in Gmail’s “smart compose” feature or Google search’s
“auto-complete” feature.
Transformer
Google’s sequence transduction model.
When trained on huge quantities of text, it can exhibit selective memory and attention to important and relevant past information.
GPT-3
Released in 2020 by OpenAI (Elon Musk’s + co.)
Generative Pre-Trained Transformers, it is a giant sequence transduction engine that learned to analyze language from a model so enormous it included almost every concept imaginable. Was trained on 45 terabytes of text, which would take 500,000 lifetimes for humans to read.
Google Brain
Release in 2021 by Google, it is a language model with 1.75 trillion parameters, which is 9x GPT-3
AlphaFold 2
Proteins are the building blocks of life, yet one aspect of proteins that have remained a mystery is how a sequence of amino acids will fold into a 3-D structure to carry out life‘s tasks.
Developed in 2020 by DeepMind, AlphaFold 2 is an AI that is trained on a large database for previously discovered 3D protein structures and is able to simulate the 3D structure of unseen proteins with similar accuracy to traditional techniques (which are expensive and takes years for each protein). Traditional methods have solved less than 0.1% of all proteins. Once a protein’s 3D structure is known, one expeditious way to discover effective treatment is to repurpose existing drugs for another ailment to see if one of them can fit into this 3D structure. Drug repurposing is quicker and will cut down costs. This also opens the door to inventing new compounds.
Quantum Computing
Quantum computing uses quantam bits, or qubits, which are typically subatomic particles such as electrons are photons. Qubits follow principles of quantum mechanics regarding how atomic and subatomic particles behave, which include unusual properties that give them super processing capabilities.
Challenges of Quantum Computing
Quantum computing is very sensitive to small disturbances in the computer and its surroundings. Even slight vibrations, electrical interference, temperature changes, or magnetic waves can cause superposition to decay or even disappear. To make a workable and scalable quantum computer, researchers have to invent new technologies and build unprecedented vacuum chambers, superconductors, and supercooling refrigerators to minimize these losses in quantum coherence, or “decoherences,” caused by environment.
GDPR
GDPR stands for a general data protection regulation.
GTPR has the vision of ultimately giving data back to the individual, so as to help people control who gets to see and use their data, and even derive value from licensing their data. It has succeeded in educating the masses on the significant risk about personal data. GDPR has required websites and apps the whole world over to rethink and Refactor their applications to minimize malicious or neglectful abuses of user data. There are large fines for a company that violates GDPR.
Some details of GDPR are not practical and in general GDPR is an impediment to AI. In its current form GDPR stipulates that companies must be transparent to people but how their data will be used. Users explicit consent for a specific purpose is needed in order for a company to start collecting that users data. Data must be protected from unauthorized use, leak, or theft. Automated decisions should be explainable and escalation to human intervention should be available upon users request.
Deep Learning
Inspired by the tangled webs of neurons in our brains, deep learning construct software layers of artificial neural networks with input and output layers. Data is fed into the input layer of the network, and a result emerges from the out of the network. In between the input and up layers maybe thousands of other layers, since the name deep learning.
