From human brain neural network to deep learning neural network, Jack from Jina AI will take you to understand “search” and “neural search” from 0 to 1.

What is neural search? How is it different from normal search? What problems does it solve? What are the advantages and disadvantages?

Watch this episode of Jina AI. Under the guidance of Jack, you will have a more comprehensive understanding of the concept, principle and similarities between Neural Search and human brain in just 6 minutes

Jack: So what is neural search?

For the full content of the video, please refer to the following text introduction.

What is a “search”

When it comes to “search”, many people’s first reaction is baidu, Google and other search engines. We input the content we want to query in the search box, and then get a series of links related to it.

In fact, search is much more than that. For example, when using the song recognition function, it is actually using recorded audio clips to search for matching songs; When you’re on Tinder, the algorithm is also searching for people it thinks you’re interested in.

Search for similar snippets of audio

In addition to these, search can do a lot of things, such as looking up answers in academic papers; Or search for a pair of shoes by searching for pictures…

Can we use any type of data to search for similar types of data? With Neural Search, it’s no longer a dream.

What is “neural search”

“Neural Search”, or Neural Search, is a concept first coined by Jina AI.

Neural Search can be understood as the abbreviation of Deep Learning for Search, which refers to using unstructured data to Search unstructured data with the help of Deep Learning technology.

Deep learning model and vector index are important components of neural search.

Neural search system includes two key elements: deep learning model and vector index, which are the differences between neural search system and traditional search system

Now, let’s use a more intuitive example to explain the concept of neural search.

This is a set of pictures of kittens and puppies. What do you notice first when you see these pictures?

First of all, you might notice, as I did, that they’re puppies and kittens; Secondly, we’ll find them pretty cute. These characteristics, while obvious, are also important.

Once you’ve got these pictures, you can sort them by their attributes

Now let’s look at the next few images. These pictures may not look so cute.

“Atypical” animal pictures

Where do these “atypical” pet pictures fit in the diagram structure? Although they belong to both cats and dogs, we thought a dimension should be added to describe their other characteristics.

Add a Y-axis to indicate how cute your pet is

Then you can put them in the right place on the diagram.

Common terms in deep learning

Now let’s look at a few common terms in deep learning.

The dimension

Represents the location of data in the graph structure. There are only two dimensions in the figure above, and they can be any number.

vector

A single data point embedded in a graph structure can be represented by a set of coordinates. In this example, there are only two dimensions (cat-dog dimension and cute-scary Dimension), so vectors can be represented by just two numbers.

In our graph structure, each pet has its own vector, and the total number is 10

The index

These vectors are pooled together and collectively referred to as “indexes”.

Definition of “index”

Once you know these terms, add some cute pet pictures. If we want to find a dog that looks like this from an animal image dataset, how do we do that?

First of all, it’s a dog and cute, so it should be in the lower right corner of the index, near other cute dogs

That’s right, we just found this puppy’s “nearest Neighbor,” which is another cute, fluffy puppy (bottom right).

Human neural networks vs virtual neural networks

This sounds simple enough, because our brains are spontaneously calling up dimensions, vectors, and indexes all the time, often without even thinking about how it works.

The human brain is a network of neurons and synapses, simply called a neural network.

Similar to human neural networks, virtual neural networks work in the same way. It builds an index based on the dimensions of the data it’s given, and it can find similar data based on its nearest neighbor, using its own neurons to search for data.

The following example is more accurate than the simple two-dimensional diagram above. Every time you eat, your brain is getting signals about the flavor of the food. Is it sweet or salty? Is it crunchy or soft?

Every food we eat is indexed into a neural network based on dimensions such as saltiness and taste; After the 3D object model is fed into the virtual neural network, it will also build indexes based on the appearance of the model, the degree of single-point concentration and so on.

Disadvantages of neural networks

All neural networks have strengths and weaknesses.

Both dolphins and squirrels have neural networks in their brains. But if we asked a dolphin to collect nuts, or a squirrel to catch fish, neither of them would do well.

Similarly, we can’t search text with a video-trained neural network, or search Chinese with a French neural network.

So, what’s Jina’s role in all this?

Jina is an open source framework that allows developers to build neural search applications for any type of data

In our next program, we’ll explore some of the key concepts behind Jina and how to get it up and running.

See you next time!