Machine Learning for Good



Artificial intelligence and machine learning are the buzzwords of the moment. Take the graph below which shows the exponential growth of submissions to the popular preprint repository ArXiv, under the header of computer science which includes machine learning, artificial intelligence, computer vision and pattern recognition. Between 2015 and 2017 alone the volume of submissions almost doubled and with companies like Google declaring themselves an AI-First company and start-ups like Magic Leap being valued at over $6 Billion, this growth doesn’t show any signs of slowing down.

From https://arxiv.org/help/stats/2017_by_area/index depicts yearly submissions to ArXiv under the header of computer science

As this growth continues, an ongoing debate is whether machine learning is useful to humans. There are quite a few misconceptions and a lot of hype- fuelled headlines. Of course, machine learning is not inherently good or bad… its contribution depends on how humans use it. There are altruistic applications such as Neural Machine Translation which helps connect people globally, regardless of their language origin. There are also examples of it being used purely for financial gain or models that at first glance appear helpful but actually contain an unsavoury bias.

By definition, machine learning is centred around digitally learning any number of arbitrary tasks without being explicitly programmed to do so. In fact, the holy grail of AI research is to finally produce an artificial general intelligence: a digital entity capable of continuous learning and development. Would such an entity be beneficial for humanity or a harbinger for its demise?

At Cookpad, we believe machine learning can be used to inspire and engage people in offline behaviours that are positive and beneficial. We’re the anti-food delivery company if you like. All food delivery companies use machine learning to capitalise on the current trend of constant connection to everyone and everything around us which belies the impression that we have very little unallocated time. So they offer the convenience of delivering a meal directly to your door. No cooking required. Ever.

Is this sustainable? At present there are hardworking individuals available to source, prepare and cook these meals. What will happen in the next decade or two when the next generation has reached adulthood by avoiding cooking, alongside any other culinary education? Who will source, prepare and cook said meals? An artificial intelligent chef?

Our hope is that we can avoid such a future. We are striving to prevent cooking being so polarising by aiding users in overcoming common barriers such as:

  • Feeling that there isn’t enough time to cook.
  • A lack of confidence or initial skill level.
  • Feeling that recipes are overwhelming.
  • Unfamiliarity with certain ingredients.
  • Not having the ability to ask for help.

And, for those that are already cooking regularly, we want you to use our platform, share your knowledge, your recipes and your experience. We want to help you become Nonnas to the world by making it easier to:

  • write down and publish your recipes.
  • recommend your recipes to those that love what you do.
  • connect with like-minded authors and users alike.
  • get the feedback that you deserve.

Much of this can and will be accomplished by good old-fashioned engineering and product design. However, we also have a number of features and product prototypes that are taking advantage of the latest developments within machine learning.

Recipe creation and representation

Our platform is fuelled by recipes. If it is cumbersome or inconvenient to create a new recipe then eventually that creation process will stagnate and cease to provide new content to our users. Similarly, if recipes are not represented well on the platform then authors will not get the level of engagement that they deserve. Examples of some of our ongoing research in this area include:

  • Smart auto cropping for food images, see the blog here.
  • Image enhancement to produce the effect that an author’s food photos were taken from a better camera. This is similar to other image enhancement work seen here.
  • Dynamic recipe text prediction and auto-complete so that creating new recipes becomes more streamlined.
  • Identifying visually clear ingredients and predicting obfuscated ones to intelligently pre-fill recipe ingredient lists.
  • Automated recipe video summarisation so that users can provide simple one-minute highlights of their recipes. This would provide an automated highlight video similar to other popular video formats.
Example of our automated image enhancement

Building a community

To begin this culinary revolution and to really make every day cooking fun we need to build and sustain a global community! More importantly, our users need to form and bond with their own individual culinary communities. Therefore, we can think of Cookpad as a global community of communities, with people sharing their ideas, feedback, life lessons and memories all through a shared passion for food. To fulfil this ambition, we are actively researching the following uses of machine learning:

  • Building a universal recipe embedding / representation so that we can compare, recommend and search across all of our content.
  • Domain specific multi language translation so that everyone around the world can experience your food.
  • How we can identify, recommend and connect users to expert authors, mentors and people with similar tastes.
  • Protecting your community by preventing the publication of inappropriate media.
An example of our recipes grouping together based on similar content

Boosting your confidence

There are many cuisines, techniques, tools and ingredients. It follows then that there should exist many different dishes, created by many different people, all of whom will have varying skill and knowledge levels. What is important to us is to be able to prioritise your preferences, skill level and knowledge base so that we can propose recipes to you as an individual. Your community is your safe zone. You should feel confident and happy with your creations. Machine learning can help us achieve this via:

  • Learning your skill level based on what you have previously cooked.
  • Learning the cooking techniques that you are comfortable with.
  • Recommending recipes that match your skill profile.
  • Adaptively increasing recipe difficulty as your confidence and skill level grows.
  • Providing guided education through selecting recipes that help you to learn a new ingredient or cooking technique.

Over the coming months, we will achieve many of these goals, release our research and demonstrate that machine learning is not just for automation but that it can be used to augment our everyday experience.

Future kitchens may well be ruled by automated robot chefs but, until then, we can strive towards creating a supportive, collaborative environment and making every day cooking fun for the people of today.

Taken from https://www.flickr.com/photos/torley/3294481634

Source: Deep Learning on Medium